BP learning algorithm based on DFP and trust region method

The Back propagation (BP) algorithm is one of the most widely used learning algorithms of Neural network (NN). The traditional BP learning algorithm has some defects, such as slow convergence rate and poor stabilization. In this paper, a new learning algorithm based on Davidon-fletcher-powell (DFP) and Trust Region method is proposed to solve these problems. Compare with other learning algorithms, DFP has advantages, such as higher searching efficiency, super-linear convergence rate and lower computation cost. On the other hand, trust region method make the new learning algorithm hold global convergence, stability, and high accuracy, especially in large residuals problems. Simulation results on XOR problem and non-linear system recognition show that this new method work well in improving the convergence rate and accuracy.